Materials for a technical, nuts-and-bolts course about increasing transparency, fairness, robustness, and security in machine learning.

• Lecture 1: Explainable Machine Learning Models
• Lecture 2: Post-hoc Explanation
• Lecture 3: Discrimination Testing and Remediation
• Lecture 4: Machine Learning Security
• Lecture 5: Machine Learning Model Debugging
• Lecture 6: Responsible Machine Learning Best Practices

Corrections or suggestions? Please file a GitHub issue.

_{^{Source: Simple Explainable Boosting Machine Example}}

• Introduction:
  • Syllabus
  • Basic Data Manipulation
  • Primer on Technical Malpractice
  • Whiteboard Notation
• Lecture Notes
• Assignment 1:
  • Model evaluation notebook
  • Full evaluations results
• Reading: Machine Learning for High-Risk Applications, Chapter 2 (pp. 33 - 50) and Chapter 6 (pp. 189 - 217)

• Python:

  • causalml
  • interpret
  • imodels
  • PiML-Toolbox
  • sklearn-expertsys
  • skope-rules
  • tensorflow/lattice

• R:

  • arules
  • elasticnet
  • gam
  • glmnet
  • quantreg
  • rpart
  • RuleFit

• Python, R or other:

  • h2o-3
  • Rudin Group code
  • xgboost

• Building from Penalized GLM to Monotonic GBM (simple)
• Building from Penalized GLM to Monotonic GBM
• Simple Explainable Boosting Machine Example
• PiML Assignment 1 Example and simple requirements.txt

• Introduction and Background:

  • An Introduction to Machine Learning Interpretability
  • Designing Inherently Interpretable Machine Learning Models
  • Psychological Foundations of Explainability and Interpretability in Artificial Intelligence
  • Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead

• Explainable Machine Learning Techniques:

  • Accurate Intelligible Models with Pairwise Interactions
  • Elements of Statistical Learning - Chapters 3,4, and 9
  • Fast Interpretable Greedy-Tree Sums (FIGS)
  • Interpretable Machine Learning - Chapter 5
  • GAMI-Net: An Explainable Neural Network Based on Generalized Additive Models with Structured Interactions
  • Neural Additive Models: Interpretable Machine Learning with Neural Nets
  • A Responsible Machine Learning Workflow with Focus on Interpretable Models, Post-hoc Explanation, and Discrimination Testing
  • This Looks Like That: Deep Learning for Interpretable Image Recognition
  • Unwrapping The Black Box of Deep ReLU Networks: Interpretability, Diagnostics, and Simplification

_{^{Source: Global and Local Explanations of a Constrained Model}}

• Lecture Notes
• Assignment 2
• Reading: Machine Learning for High-Risk Applications, Chapter 2 (pp. 50 - 80) and Chapter 6 (pp. 208 - 230)

• Python:

  • allennlp
  • alibi
  • anchor
  • DiCE
  • interpret
  • lime
  • shap
  • PiML-Toolbox
  • tf-explain

• R:

  • ALEPlot
  • DALEX
  • ICEbox
  • iml
  • Model Oriented
  • pdp
  • shapFlex
  • vip

• Python, R or other:

  • h2o-3

• Global and Local Explanations of a Constrained Model
• Building from Penalized GLM to Monotonic GBM
• Monotonic XGBoost models, partial dependence, individual conditional expectation plots, and Shapley explanations
• Decision tree surrogates, LOCO, and ensembles of explanations

• Introduction and Background:

  • On the Art and Science of Explainable Machine Learning
  • Proposed Guidelines for the Responsible Use of Explainable Machine Learning

• Post-hoc Explanation Techniques:

  • A Unified Approach to Interpreting Model Predictions
  • Anchors: High-Precision Model-Agnostic Explanations
  • Elements of Statistical Learning - Section 10.13
  • Extracting Tree-Structured Representations of Trained Networks
  • Interpretability via Model Extraction
  • Interpretable Machine Learning - Chapter 6 and Chapter 7
  • Peeking Inside the Black Box: Visualizing Statistical Learning with Plots of Individual Conditional Expectation
  • Towards Better Understanding of Gradient-based Attribution Methods for Deep Neural Networks
  • Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models
  • “Why Should I Trust You?” Explaining the Predictions of Any Classifier

• Problems with Post-hoc Explanation:

  • General Pitfalls of Model-Agnostic Interpretation Methods
  • Limitations of Interpretable Machine Learning Methods
  • When Not to Trust Your Explanations

_{^{Source: Lecture 3 Notes}}

• Lecture Notes
• Software Example: Testing a Constrained Model for Discrimination and Remediating Discovered Discrimination
• Assignment 3

Python:

• aequitas
• AIF360
• Themis

• Introduction and Background:

  • Fairness and Machine Learning - Introduction
  • NIST SP1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence
  • Fairness Through Awareness

• Discrimination Testing and Remediation Techniques:

  • Certifying and Removing Disparate Impact
  • Data Preprocessing Techniques for Classification Without Discrimination
  • Decision Theory for Discrimination-aware Classification
  • Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification Without Disparate Mistreatment
  • Learning Fair Representations
  • Mitigating Unwanted Biases with Adversarial Learning

_{^{Source: Responsible Machine Learning}}

• Lecture Notes
• Software Example: Attacking a Machine Learning Model
• Assignment 4

Python:

• cleverhans
• foolbox
• robustness

• Introduction and Background:

  • A Marauder’s Map of Security and Privacy in Machine Learning
  • BIML Interactive Machine Learning Risk Framework
  • The Security of Machine Learning
  • Proposals for model vulnerability and security

• Machine Learning Attacks:

  • Membership Inference Attacks Against Machine Learning Models
  • Stealing Machine Learning Models via Prediction APIs
  • Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures
  • Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers

• Links from Lecture 4:

  • Mitre Adversarial Threat Matrix
  • Robust ML
  • ISIS 'still evading detection on Facebook', report says
  • Researchers bypass airport and payment facial recognition systems using masks
  • Slight Street Sign Modifications Can Completely Fool Machine Learning Algorithms

_{^{Source: Real-World Strategies for Model Debugging}}

• Lecture Notes
• Software Example: Debugging a Machine Learning Model
• Assignment 5

• DALEX (R)
• SALib (Python)
• themis-ml (Python)
• What-if Tool (Browser)

• Introduction and Background:

  • Debugging Machine Learning Models
  • Why you should care about debugging machine learning models
  • Real-World Strategies for Model Debugging

• Links from Lecture 5:

  • AI Incident Tracker
  • AI Incident Database
  • Debugging Machine Learning Models
  • Underspecification Presents Challenges for Credibility in Modern Machine Learning

_{^{A Responsible Machine Learning Workflow Diagram. Source: Information, 11(3) (March 2020).}}

• Lecture Notes
• Assignment 6 (Final Assessment)

• Awesome Machine Learning Interpretability

• Introduction and Background:

  • NIST AI Risk Management Framework
  • Interagency Guidance on Model Risk Management (SR 11-7)
  • Eight Principles of Responsible Machine Learning
  • Principles for Accountable Algorithms and a Social Impact Statement for Algorithms
  • Responsible AI Practices

• Links from Lecture 6:

  • Predicting Good Probabilities With Supervised Learning